Neural Network Based Approach for Short-Term Load Forecasting

muscleblouseΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 10 μήνες)

72 εμφανίσεις

Neural Network Based Approach
for Short
-
Term Load Forecasting

Zainab

H. Osman,

Mohamed L.
Awad
,

Tawfik

K. Mahmoud


Power Systems Conference and
Exposition, 2009, IEEE/PES

1

Outline


Intro


Load analysis


Artificial Neural Network(ANN) structure


Input variables


Training data


Network topology


Result


Conclusion

2

Introduction


This paper analyze the relationship between factors
and electricity load



By the analytical results, they choose the most
correlated factors in different seasons to feed in the
artificial neural networks(ANN).

3

Load analysis


Different characteristic of the power system and it’s
load pattern significantly affect the ANN model. It is
important to extract load characteristic such as
periodicity and trends before design.




4

Load analysis


Characteristic of first week in January, April, July,
October, represent four seasons

5

Load analysis

6

Load analysis


Correlation between load demand and weather parameters:
temperature, dew point, wind speed, humidity

7

Fall

sprin
g

ANN structure


input variables


By previous load analysis, the
historical load
are the
most correlated parameter to the forecasted
load.



Temperature

are highly related to load demand in
summer and
spring.
Temperature

and
humidity

seem
to be the most affecting weather parameter. In winter
and fall wind speed effect can be negligible.



Since we forecast the hourly load, the variables are
hourly values.

8

ANN
structure


training data


Training is the process to determine the ANN’s weights and
biases. The training data should cover a wide range of input
patterns sufficient enough to train the network



Typically, ANNs are trained following a supervised pattern,
the
desired output is given for each input and the training
process then adjusts the weights and biases to match the
desired output

9

ANN structure


training data


In this paper, minimum distance between forecasted
input variables and desire outcome is calculate for the
entire database.



Data who does not achieve the condition will be filtered,
in order to eliminate odd data and sudden load change
due to drastic weather changes,

condition :

𝑃
𝑑
,
𝑡

𝑃
^
(
𝑑
,
𝑡
)
<
𝑎
𝑖

10

ANN structure


training data


The load pattern is divided into seven patterns
represent 7 days of a week. And each season has its
own training vectors



The training information is select on similar weather
condition days of the forecasted day.



For each season, former 60 days of data is used for
training, and the latter 30 days is for testing.

11

ANN structure


network topology


A three
-
layer feed forward neural network is used

Output layer

Hidden layer

Input layer

12

ANN structure


network topology


The approach of selecting proper number of hidden
neurons is:

1.
Set the estimated optimal number of neurons as square root
of the product of number of inputs times number of outputs.

2.
Increment by one to find the minimum forecast error



13

ANN structure


network topology


Inputs of ANN


common


Load value of
previous hour


The load value of
one, two and three days preceding

the forecasted day
at the same and the previous hour.


The load value of the same and previous hour in
previous week


The forecasted
hourly

temperature, dew point, relative humidity and
wind speed


Summer and Spring : previous hour’s
temperature, dew point,
relative humidity and wind speed


Winter and Fall : only previous hour’s temperature




14

Result


2004 Egypt Unified System load, which now use
regression/trending method for short
-
term load forecast.



Since the neural network work well on weekdays, but
according to the results, it is not enough to forecast
weekends and holidays

15

Result


Include/exclude the effect of weekends and holidays


Proposed method







Conventional regression method

16

Conclusion


This paper use a threshold to eliminate special events
in training procedure. It can be combined with other
holiday and weekend forecast method.



Analysis correlation coefficients of input variables and
electricity loads is a good way to decide which factors
should be put in.



17